,
9 tweets,
4 min read
Read on Twitter
@yudapearl You've redefined credibility so that it ignores everything learned in applied economics since 1990. The need to focus on "natural experiments" to construct "t-DAGs" is not a theoretical conclusion, but an empirical one based on trying different methods in thousands of cases.
@yudapearl Nothing you have said distinguishes between the (disastrous) approach of empirical researchers in the pre-1990 period, analogous to estimating models based on large DAGs with dozens of exclusion restrictions, and the enormously fruitful modern approach where
@yudapearl inference in all papers, whether with explicit structural models or not, rests on well-understood t-DAGs like IV, diff-in-diff or RD, where a well-developed theory allows for *transparent* robustness testing.
@yudapearl You're like a string theorist trying to tell engineers building a rocket to the moon that they need to focus their efforts on mastering string theory before they apply Newtonian gravity because string theory provides the foundational explanation and justification for gravity.
@yudapearl Do structural models provide the philosophical basis for counterfactual inference? Yes. Is it possible to write informative empirical papers without an explicit structural model? Absolutely. Can a structural model aid in assessing questions like external validity? Of course.
@yudapearl Because you have limited familiarity with applied work, you fail to appreciate the lessons of the credibility revolution and you advocate a blinkered and antiquated view of causal inference with a sound theoretical basis but disastrous practical recommendations.
@yudapearl Instead of dismissing Angrist, learn something from him. Read his empirical papers. Then write textbooks showing how DAGs+Angrist type techniques are complementary and together can rule the galaxy as father and son.
@yudapearl @MariaGlymour do you agree that epidemiologists would be well-served paying more attention to searching hard for clean identification rather than writing down large DAGs and doing "automated Angrist" by finding IVs with DAGgity? (not rhetorical, curious what practitioners think)
@yudapearl @MariaGlymour My view is that the "t-DAG"--the kernel of identification such as a natural experiment or RD needs to come first--then one can play with the full DAG to assess threats to identification in DAGgity. Have you ever "discovered" your identification strategy in DAGgity?
